ToolGrad: Efficient Tool-Use Dataset Generation with Textual Gradients¶
Local core-mechanism evidence. Paper outcomes and local results are separate.
Paper information¶
| Field | Value |
|---|---|
| Paper | ToolGrad: Efficient Tool-Use Dataset Generation with Textual Gradients |
| Affiliations | Google, first author's affiliation |
| First public release | 2025-08-06 |
| Authors' code | Authors' code. |
| Adapter / method | toolgrad |
| Local code | src/auto_research/agent_research/latest_20260914.py |
Original paper¶
Background and contribution¶
Work backward from a target answer to propose tool traces, then use textual gradients to locate and revise failed calls, reducing manual annotation.
flowchart LR
A[Target answer] --> T[Proposed tool trace] --> F[Failure feedback] --> E[Textual edits and deduplication]Key figure from the paper¶
Original Figure 2: ToolGrad. From the paper; copyright belongs to the authors. Not a local result.
Core mechanism¶
The deterministic reference edits failing calls, deduplicates traces, and exposes selection and state statistics for boundary tests.
Paper results¶
ToolGrad-12B scores 83.1 on BFCL versus Gemini 2.5 Pro's 83.2, as reported in the paper. No local benchmark or inferred online gain.
Local reproduction¶
Three-seed metrics use matched fixtures and budgets. The deterministic reference exposes edit/dedup behavior and tests its boundaries.
Reproduction limits¶
L1 answer-first and deduplicated textual edits only. No 12B training or BFCL run.